Problem calling MPI with GPU arrays (use_device_ptr, use_device_addr)

[caplanr]

Hi,

I am trying to compile and run our code (github.com/predsci/pot3d) on Stampede3 on the MAX 1550 GPUs.

The code uses "GPU-aware" MPI calls using OpenMP target's "use_device_ptr()".

(This works with the NVIDIA compiler with the HPC-X MPI library across NVIDA GPUs).

On the Intel GPUs, I am getting a seg fault at the MPI calls.

The compiler mentions that the "use_device_ptr()" can only be used with C pointers in Fortran, so I switched them to "use_device_addr()", but that also seg faults.

Are GPU-aware MPI calls supported in ifx+impi (I am using 2025.0.0)?

If so, what do I need to do to get it to work?

Can you clarify the difference between use_device_ptr() and use_device_addr()?

From here, it seems the difference is not clear?

https://forums.developer.nvidia.com/t/omp-target-data-use-device-ptr-vs-use-device-addr/290317

Thanks!

 - Ron

[TobiasK · Accepted solution]

@caplanr when passing GPU pointer to Intel MPI you have to set
export I_MPI_OFFLOAD=1
before running the application

[caplanr]

Thanks!

It is now working!

Another tip for someone with the same issue:

I also had to set:

export LIBOMPTARGET_DEVICES=SUBDEVICE

in order for the MPI to see all 8 tiles of the 4 MAX 1550 GPUs.

The default was only seeing 1 GPU device (on Stampede3).

 - Ron